Research And Application Of Wood Surface Defect Detection Based On Deep Learning

In the process of wood processing and production,various defects often exist on the surface of wood,such as dead knots,live knots,bark,cracks,and gaps,which directly affect the quality and processing efficiency of wood.Therefore,wood surface defect detection is an important aspect in production.Currently,many wood processing companies mainly rely on manual methods for wood defect detection,which is inefficient and imprecise.This project takes deep learning technology as a starting point to achieve automatic detection of wood defects,which is of great significance for improving the intelligent level of China’s wood processing industry.This project is based on object detection technology to research and design a system for wood surface defect detection and sorting,aiming to improve the efficiency and accuracy of defect detection and replace manual sorting tasks.The main research contents of this project are as follows:1.Introduce the background of wood defect detection,elucidate the research purpose and significance of this project.Provide a detailed introduction to the current research status of traditional wood defect detection methods and deep learning technology in the field of wood defect detection both domestically and internationally,and finally define the research content of this project.2.Wood data collection and preprocessing.Firstly,the data of Eucalyptus veneer is collected through the image acquisition unit,and the dataset is preprocessed using image processing techniques,including target region cropping,exposure processing,etc.Then,traditional data augmentation methods and class coverage methods are used to expand the dataset.Finally,the dataset is divided to provide a data foundation for subsequent research work.3.Wood defect detection model construction.This project proposes a CBi2-YOLO model based on the YOLOv5 algorithm to achieve wood defect detection.To address the issues of defect recognition accuracy and real-time performance,the main backbone network of the model adopts the C2 F structure to reduce computational complexity,while the bottleneck layer network uses the BiFPN structure for multi-scale feature fusion,enabling it to quickly and accurately extract wood surface defect features.Finally,a tiny object detection head is added to improve the accuracy of small object defect detection.The system achieves defect size selection by calculating the defect area size through image thresholding.4.System application verification.The effectiveness of the proposed method in this study is validated through ablation experiments and comparative experiments,and the accuracy of the system sorting is tested in actual production environment to ultimately verify the feasibility of the entire system.Based on multiple experimental results,the proposed method in this study achieved a mAP@0.5 value of 0.727 in wood defect detection,which is about 10% higher than the original structure.Through comparative experiments,this method can meet the requirements of system detection accuracy and real-time performance.The entire wood surface defect detection system was tested with around 2,100 Eucalyptus veneers,and the system sorting accuracy was approximately 97.7%.The system has been tested and proven to be able to replace manual labor.